This spring, our consortium partner Eticas Foundation published its Guide to Community-Led AI Audits 2026 — a flagship report that brings together six real-world audits into a comprehensive, freely available methodology for holding AI systems accountable.
The guide builds on work first developed with the support of the DIVERSIFAIR project. An earlier version was
shared with DIVERSIFAIR partners last November. This new edition is a fully revised, publicly released
resource, designed to be used by civil society organisations, researchers, advocates, and communities
around the world.

From adversarial auditing to community-led practice

The 2026 guide is the natural successor to Eticas Foundation’s 2023 Adversarial Algorithmic Auditing Guide, which introduced the concept of auditing AI systems without institutional cooperation — from the outside, using only what can be observed, tested, or reconstructed. Where that guide laid out the theoretical foundations, the 2026 edition is grounded in years of practice.
Algorithmic and AI systems now shape some of the most consequential decisions in people’s lives: who receives welfare benefits, how parole risk is assessed, how much someone pays for a taxi ride or health insurance. Yet the people most directly affected by these systems are rarely involved in evaluating them — and often have no idea the systems exist.
Community-led auditing takes a different approach. It treats the people most impacted by an AI system as essential to understanding how it actually works. Communities help define the research questions, identify which outcomes matter, contribute to data collection, and interpret findings. The result is an evidence base grounded in lived experience, not just technical documentation.

Six audits across four countries

The 2026 guide documents six community-led audits conducted between 2018 and 2024, each carried out without institutional cooperation. Together, they show what community-led auditing can achieve — and what it consistently uncovers.

• RisCanvi: Catalonia’s parole risk algorithm had been shaping the futures of thousands of incarcerated people for 15 years. Eticas partnered with Iridia to conduct the first independent audit of the system, finding no statistically significant relationship between its 43 risk factors and its outcomes. The scores were effectively random. (Find out more in this dedicated article).
• Serbia’s Social Card Registry: Introduced to make welfare distribution fairer, the system was found to exclude 47% of assessed cases wrongfully. Around 20% of beneficiaries were cut off without a written decision, removing any right of appeal. Eticas partnered with A11, whose constitutional challenge before Serbia’s highest court was directly supported by the audit findings
• Facial recognition and people with Down Syndrome: Eticas partnered with Cedown Jerez to test two systems with 40 participants. Azul, used by Zurich Insurance to set premiums, classified half of women with Down Syndrome as children. DeepFace correctly identified the gender of all male participants but fewer than half of female ones.
• Uber, Bolt, and Cabify pricing: — Following complaints to Spain’s competition authority, Eticas partnered with Taxi Project 2.0 and Observatorio TAS to monitor pricing across 15 routes over three months. The three platforms showed a statistically significant correlation in fares, and customers in lower-income neighbourhoods paid more for equivalent trips.
• Uber’s service in Roma Madrid neighbourhoods: In a follow-up study, Eticas found that Uber was unavailable for 27% of trip requests from Roma neighbourhoods. In non-Roma areas, rides were always available. When service was provided, wait times were 1.4 times longer.

What the guide provides

The 2026 guide offers a practical 10-step framework covering the full audit process — from choosing which system to investigate, to building community partnerships, designing methodology, analysing findings, and producing recommendations for specific audiences. It also includes a taxonomy of algorithmic and AI systems, a catalogue of bias types, and an overview of the research methods used across the six audits.

All underlying data, code, and quantitative analyses are publicly available and openly documented, so findings can be independently verified. The full repository is available at github.com/Eticas-Foundation.